Model generation method, speech recognition method, device, medium, and apparatus

Abstract

The present disclosure relates to a model generation method, a speech recognition method, an apparatus, a medium and a device. The method comprises: obtaining a plurality of first labeled audio samples and a plurality of unlabeled audio samples in a target domain; performing self-supervised model pre-training on a first initial model using the plurality of unlabeled audio samples to obtain a pre-trained model; generating a first pseudo label for each unlabeled audio sample using the plurality of first labeled audio samples and the pre-trained model; and performing supervised fine-tuning training on the pre-trained model using the first pseudo labels and the plurality of unlabeled audio samples to obtain a target speech recognition model. In this way, the pseudo label can compensate for the problem of low model accuracy caused by insufficient labeled audio samples in the target domain, and improve the accuracy of the speech recognition model. In addition, the general domain training data can also avoid interfering with the speech recognition ability of the model in the target domain, and ensure the speech recognition ability of the model in the target domain.

Description

Technical Field

[0001] This disclosure relates to the field of speech recognition technology, specifically to a model generation method, a speech recognition method, an apparatus, a medium, and a device. Background Technology

[0002] The speech recognition performance of a speech recognition model is closely related to the data used for its training. Speech recognition models usually have certain requirements for the scale of labeled data in the target domain to which they are applied. When there is insufficient labeled data in the target domain, there are usually two ways to handle this: (1) First, perform self-supervised pre-training based on a large amount of unlabeled data in the target domain, and then perform supervised fine-tuning using labeled data in the target domain. However, there is usually a small amount of labeled data in the target domain, so the accuracy of the speech recognition model cannot be guaranteed; (2) Use existing labeled data in the general domain to pre-train the model, and then perform supervised fine-tuning using labeled data in the target domain. However, the labeled data in the general domain may not match the target domain of the final application, which may interfere with the speech recognition ability of the speech recognition model in the target domain, thus affecting the speech recognition ability of the speech recognition model in the target domain. Summary of the Invention

[0003] This section is provided to briefly introduce the concepts, which will be described in detail in the Detailed Description section later. This section is not intended to identify key or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0004] In a first aspect, this disclosure provides a method for generating a speech recognition model, including:

[0005] Obtain multiple first-labeled audio samples and multiple unlabeled audio samples within the target domain;

[0006] The first initial model is pre-trained using the multiple unlabeled audio samples to obtain a pre-trained model.

[0007] Using the plurality of first labeled audio samples and the pre-trained model, a first pseudo-label is generated for each of the unlabeled audio samples;

[0008] Using each of the first pseudo-labels and the plurality of unlabeled audio samples, the pre-trained model is subjected to supervised fine-tuning training to obtain the target speech recognition model.

[0009] Secondly, this disclosure provides a speech recognition method, the method comprising:

[0010] Obtain the audio data to be recognized;

[0011] The audio data is input into a pre-trained target speech recognition model to obtain the target text corresponding to the target audio data, wherein the target speech recognition model is trained according to the speech recognition model generation method provided in the first aspect of this disclosure.

[0012] Thirdly, this disclosure provides a speech recognition model generation apparatus, comprising:

[0013] The first acquisition module is used to acquire multiple first labeled audio samples and multiple unlabeled audio samples within the target domain;

[0014] The first pre-training module is used to perform self-supervised model pre-training on the first initial model using the multiple unlabeled audio samples to obtain a pre-trained model.

[0015] The generation module is used to generate a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model;

[0016] The fine-tuning module is used to perform supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples to obtain the target speech recognition model.

[0017] Fourthly, this disclosure provides a voice recognition device, the device comprising:

[0018] The second acquisition module is used to acquire the audio data to be recognized;

[0019] A speech recognition module is used to input the audio data into a pre-trained target speech recognition model to obtain the target text corresponding to the target audio data, wherein the target speech recognition model is trained according to the speech recognition model generation method provided in the first aspect of this disclosure.

[0020] Fifthly, this disclosure provides a computer-readable medium having a computer program stored thereon, which, when executed by a processing device, implements the steps of the method provided in the first or second aspect of this disclosure.

[0021] Sixthly, this disclosure provides an electronic device, comprising:

[0022] A storage device having at least one computer program stored thereon;

[0023] At least one processing device is configured to execute the at least one computer program in the storage device to implement the steps of the method provided in the first or second aspect of this disclosure.

[0024] In the above technical solution, the model is first pre-trained using unlabeled audio samples within the target domain. Then, pseudo-labels for the unlabeled audio samples within the target domain are generated using labeled audio samples and the pre-trained model. Finally, the pre-trained model is fine-tuned using the unlabeled audio samples and the corresponding pseudo-labels to obtain the target speech recognition model. This allows pseudo-labels to compensate for the low accuracy of the speech recognition model caused by insufficient labeled audio samples in the target domain, thereby improving the accuracy of the speech recognition model. Furthermore, the entire model generation stage uses training data within the target domain, avoiding interference from general domain training data on the speech recognition model's ability in the target domain, thus ensuring the target speech recognition model's speech recognition capability within the target domain. In addition, the accuracy of the speech recognition model can be guaranteed with only a small number of labeled audio samples, eliminating the need for extensive manual annotation, saving manpower and improving the training efficiency of the speech recognition model.

[0025] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0026] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale. In the drawings:

[0027] Figure 1 This is a flowchart illustrating a speech recognition model generation method according to an exemplary embodiment.

[0028] Figure 2 This is a schematic diagram illustrating a speech recognition model generation method according to an exemplary embodiment.

[0029] Figure 3 This is a flowchart illustrating a speech recognition model generation method according to another exemplary embodiment.

[0030] Figure 4 This is a schematic diagram illustrating a speech recognition model generation method according to another exemplary embodiment.

[0031] Figure 5 This is a block diagram illustrating a speech recognition model generation apparatus according to an exemplary embodiment.

[0032] Figure 6 This is a block diagram illustrating a speech recognition device according to an exemplary embodiment.

[0033] Figure 7 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0034] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0035] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0036] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0037] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0038] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0039] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0040] All actions involving the acquisition of signals, information, or data in this disclosure are carried out in accordance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0041] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0042] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution, based on the prompt message.

[0043] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0044] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0045] Meanwhile, it is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0046] Figure 1 This is a flowchart illustrating a speech recognition model generation method according to an exemplary embodiment. For example... Figure 1 As shown, the method may include the following S101 to S104.

[0047] In S101, multiple first labeled audio samples and multiple unlabeled audio samples are obtained within the target domain.

[0048] In this disclosure, the term "domain" refers broadly to certain characteristics of the audio data itself, such as the speaking context (noisy or quiet background), speaking style (reading aloud, natural conversation, tone, accent, etc.), audio quality (sampling rate, near and far fields, compression methods, etc.), and the topic of the speaking content (casual conversation, professional field, etc.). The space containing the general audio samples used to train the speech recognition model is called the source domain, and the space containing the audio samples in a specific scenario is called the target domain. Data within the target domain is referred to as "in-domain" (ID) data, and data within the source domain is referred to as "out-of-domain" (OOD) data.

[0049] The first labeled audio sample includes the first audio sample within the target domain and the labeled text corresponding to the first audio sample.

[0050] For example, the above fields are professional fields, and the target fields are fields such as medicine and aerospace.

[0051] In S102, the first initial model is pre-trained using multiple unlabeled audio samples to obtain a pre-trained model.

[0052] In this disclosure, the first initial model can be a traditional acoustic model + language model structure, or it can be an end-to-end model structure. This disclosure does not specifically limit the specific model used. The training objective of the first initial model can be set according to the actual task requirements. After pre-training is complete, the pre-trained model can be obtained.

[0053] In S103, multiple first labeled audio samples and a pre-trained model are used to generate the first pseudo-label for each unlabeled audio sample.

[0054] In this disclosure, the first pseudo-label is the pseudo-text corresponding to the unlabeled audio sample.

[0055] In S104, the pre-trained model is supervised fine-tuned using each first pseudo-label and multiple unlabeled audio samples to obtain the target speech recognition model.

[0056] Specifically, unlabeled audio samples can be used as input to a pre-trained model, and the first pseudo-label corresponding to the unlabeled audio sample can be used as the target output of the pre-trained model for supervised fine-tuning training, so as to fine-tune the model parameters and obtain the target speech recognition model.

[0057] In the above technical solution, the model is first pre-trained using unlabeled audio samples within the target domain. Then, pseudo-labels for the unlabeled audio samples within the target domain are generated using labeled audio samples and the pre-trained model. Finally, the pre-trained model is fine-tuned using the unlabeled audio samples and the corresponding pseudo-labels to obtain the target speech recognition model. This allows pseudo-labels to compensate for the low accuracy of the speech recognition model caused by insufficient labeled audio samples in the target domain, thereby improving the accuracy of the speech recognition model. Furthermore, the entire model generation stage uses training data within the target domain, avoiding interference from general domain training data on the speech recognition model's ability in the target domain, thus ensuring the target speech recognition model's speech recognition capability within the target domain. In addition, the accuracy of the speech recognition model can be guaranteed with only a small number of labeled audio samples, eliminating the need for extensive manual annotation, saving manpower and improving the training efficiency of the speech recognition model.

[0058] The following is a detailed explanation of the specific implementation method for generating the first pseudo-label for each unlabeled audio sample using multiple first labeled audio samples and a pre-trained model in S103 above. Specifically, it can be achieved through the following steps (1) and (2):

[0059] Step (1): Supervised fine-tuning training of the pre-trained model is performed using multiple first labeled audio samples to obtain the first speech recognition model.

[0060] Specifically, the first audio sample in the first labeled audio sample can be used as the input of the pre-trained model, and the labeled text corresponding to the first audio sample can be used as the target output of the pre-trained model. Supervised fine-tuning training of the model can be performed to fine-tune the model parameters, thereby obtaining the first speech recognition model.

[0061] Step (2): For each unlabeled audio sample, perform speech recognition on the unlabeled audio sample using the first speech recognition model to obtain the first pseudo-label of the unlabeled audio sample.

[0062] The following is combined with Figure 2 Provide a detailed explanation of the specific steps involved in training a speech recognition model. For example... Figure 2 As shown, this can be achieved through steps ① to ④:

[0063] Step ①: Using multiple unlabeled audio samples within the target domain, perform self-supervised model pre-training on the first initial model to obtain the pre-trained model;

[0064] Step 2: Supervised fine-tuning training of the pre-trained model is performed using multiple first labeled audio samples within the target domain to obtain the first speech recognition model;

[0065] Step 3: Use the first speech recognition model to decode each unlabeled audio sample in the target domain, that is, use the first speech recognition model to perform speech recognition on each unlabeled audio sample to obtain the first pseudo-label of each unlabeled audio sample.

[0066] Step 4: Using each unlabeled audio sample and the corresponding first pseudo-label, perform supervised fine-tuning training on the pre-trained model to obtain the target speech recognition model.

[0067] Figure 3 This is a flowchart illustrating a speech recognition model generation method according to another exemplary embodiment. For example... Figure 3 As shown, before S104 above, the above method may also include the following S105 to S107.

[0068] In S105, multiple second-labeled audio samples are acquired from the source domain.

[0069] In this disclosure, the second labeled audio sample includes a second audio sample within the source domain and the corresponding labeled text of the second audio sample.

[0070] In S106, supervised model pre-training is performed on N second initial models using multiple second labeled audio samples to obtain N second speech recognition models.

[0071] In this disclosure, for each of the N second initial models, a supervised model pre-training can be performed by taking a second audio sample from the second labeled audio samples as the input of the second initial model and taking the labeled text corresponding to the second audio sample as the target output of the second initial model, so as to obtain the second speech recognition model corresponding to the second initial model.

[0072] Additionally, N ≥ 1. In one implementation, N = 1, meaning that a second initial model can be pre-trained using multiple second labeled audio samples to obtain a second speech recognition model.

[0073] In another implementation, N > 1, that is, for each output model in the N second initial models, supervised model pre-training can be performed on the second initial model using multiple second labeled audio samples to obtain the second speech recognition model corresponding to the second initial model. Thus, N second speech recognition models can be obtained.

[0074] Each of the second initial models can be a traditional acoustic model + language model structure, or it can be an end-to-end model structure. This disclosure does not specifically limit the specific structure. Furthermore, the structures of the N second initial models can be different from each other, and they may or may not include models with the same structure as the first initial model. That is, the N second initial models and the first initial model are different from each other.

[0075] Preferably, the N second initial models are all different from each other and have significant structural differences. This allows for the simultaneous training of multiple second speech recognition models with different structures and sizes, thereby enhancing the diversity and complementarity of the second pseudo-labels, improving the accuracy of the target pseudo-labels, and ultimately improving the accuracy of the target speech recognition model.

[0076] return Figure 3 In S107, for each of the N second speech recognition models, speech recognition is performed on each unlabeled audio sample using the second speech recognition model to obtain multiple second pseudo-labels for the unlabeled audio sample corresponding to the second speech recognition model.

[0077] At this point, S104 above can utilize each first pseudo-label, multiple second pseudo-labels corresponding to each second speech recognition model, and multiple unlabeled audio samples to perform supervised fine-tuning training on the pre-trained model, thereby obtaining the target speech recognition model. Here, the second pseudo-label is the pseudo-text of the corresponding unlabeled audio sample.

[0078] Specifically, this can be achieved through the following steps [1] and [2]:

[0079] Step [1]: Select multiple target pseudo-labels from the multiple second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model.

[0080] Step [2]: Use multiple target pseudo-labels and the unlabeled audio samples corresponding to each target pseudo-label to perform supervised fine-tuning training on the pre-trained model to obtain the target speech recognition model.

[0081] Specifically, the unlabeled audio samples corresponding to the target pseudo-label can be used as the input to the pre-trained model, and the target pseudo-label can be used as the target output of the pre-trained model to perform supervised fine-tuning training of the model parameters, thereby obtaining the target speech recognition model.

[0082] In the above implementation, when the number of first labeled audio samples in the target domain is small, the recognition accuracy of the first speech recognition model obtained by supervising the pre-trained model using these samples may be low. Consequently, the accuracy of the first pseudo-labels for unlabeled audio samples generated by the first speech recognition model may also be low. Simultaneously, when the source domain differs significantly from the target domain, the N second speech recognition models obtained by supervising the pre-training of multiple second labeled audio samples in the source domain may not be highly accurate in recognizing unlabeled audio samples in the target domain. Consequently, the recognition accuracy of the second pseudo-labels for unlabeled audio samples generated by each first speech recognition model may also be low. Therefore, it is necessary to first select target pseudo-labels with higher accuracy from the multiple first pseudo-labels generated by the first speech recognition model and the multiple second pseudo-labels generated by each second speech recognition model. Then, the pre-trained model is supervised and fine-tuned based on these target pseudo-labels to obtain a target speech recognition model with higher recognition accuracy.

[0083] The following is a detailed description of the specific implementation method for selecting multiple target pseudo-labels from multiple second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model in the above step [1].

[0084] In one implementation, N=1. In this case, for each unlabeled audio sample, the first similarity between the first pseudo-label and the second pseudo-label of the unlabeled audio sample can be calculated. If the first similarity is greater than the first preset similarity threshold, the target pseudo-label is determined from the first pseudo-label and the second pseudo-label of the unlabeled audio sample.

[0085] Specifically, if the first similarity is greater than the first preset similarity threshold, it indicates that the consistency between the first pseudo-label and the second pseudo-label of the unlabeled audio sample is high, that is, the accuracy of the first pseudo-label and the second pseudo-label of the unlabeled audio sample is high. In this case, the target pseudo-label can be determined from the first pseudo-label and the second pseudo-label of the unlabeled audio sample for fine-tuning of the pre-trained model. If the first similarity is less than or equal to the first preset similarity threshold, it indicates that the consistency between the first pseudo-label and the second pseudo-label of the unlabeled audio sample is poor, that is, the accuracy of the first pseudo-label and the second pseudo-label of the unlabeled audio sample cannot be guaranteed and cannot be used for fine-tuning of the pre-trained model. In this case, the first pseudo-label and the second pseudo-label of the unlabeled audio sample are discarded.

[0086] In another implementation, N>1, then for each unlabeled audio sample, the second similarity between each pair of the first pseudo-label and N second pseudo-labels of the unlabeled audio sample can be calculated; if the maximum value of the second similarity between each pair of the unlabeled audio sample is greater than the second preset similarity threshold, then the target pseudo-label is determined from the two pseudo-labels corresponding to the maximum value.

[0087] Specifically, if the maximum value of the second similarity between any two pairs of unlabeled audio samples is greater than the second preset similarity threshold, it indicates that the consistency between the two pseudo-labels corresponding to the maximum value is high, meaning that the accuracy of the maximum value and the two corresponding pseudo-labels of the unlabeled audio sample is high. In this case, the target pseudo-label can be determined from the maximum value and the two corresponding pseudo-labels of the unlabeled audio sample for fine-tuning the pre-trained model. If the maximum value of the second similarity between any two pairs of unlabeled audio samples is less than or equal to the second preset similarity threshold, it indicates that the consistency between any two pairs of the first pseudo-label and the N second pseudo-labels of the unlabeled audio sample is poor, meaning that the accuracy of the first pseudo-label and the N second pseudo-labels of the unlabeled audio sample cannot be guaranteed and cannot be used for fine-tuning the pre-trained model. In this case, the first pseudo-label and the N second pseudo-labels of the unlabeled audio sample are discarded.

[0088] In addition, in this disclosure, the similarity between two pseudo-labels can be measured by using cosine distance, word error rate between two pseudo-labels, Euclidean distance, etc. The embodiments of this disclosure do not specifically limit the calculation method of the first similarity and the second similarity.

[0089] The following provides a detailed description of the specific implementation method for determining the target pseudo-label from the first and second pseudo-labels of the unlabeled audio sample. Specifically, this can be achieved through various implementation methods. In one implementation method, the first pseudo-label of the unlabeled audio sample can be determined as the target pseudo-label.

[0090] In another implementation, the second pseudo-label of the unlabeled audio sample can be identified as the target pseudo-label.

[0091] In another implementation, one of the first and second pseudo-labels of the unlabeled audio sample can be randomly selected as the target pseudo-label. This increases the diversity and complementarity of the target pseudo-labels, improves their accuracy, and consequently enhances the accuracy of the target speech recognition model.

[0092] In another implementation, the pseudo-label generated by a speech recognition model with a low word error rate from the first and second pseudo-labels of the unlabeled audio sample is determined as the target pseudo-label. This improves the accuracy of the target pseudo-label, thereby improving the accuracy of the target speech recognition model.

[0093] Alternatively, a similar method to that used above to determine the target pseudo-label from the first and second pseudo-labels of the unlabeled audio sample can be used to determine the target pseudo-label from the two pseudo-labels corresponding to the maximum value, which will not be elaborated further in this disclosure.

[0094] The following is combined with Figure 4 Please provide a detailed explanation of the specific methods for training the speech recognition model. For example... Figure 4 As shown, this can be achieved through steps ① to ⑦:

[0095] Step ①: Using multiple unlabeled audio samples within the target domain, perform self-supervised model pre-training on the first initial model to obtain the pre-trained model;

[0096] Step 2: Supervised fine-tuning training of the pre-trained model is performed using multiple first labeled audio samples within the target domain to obtain the first speech recognition model;

[0097] Step 3: Use the first speech recognition model to decode each unlabeled audio sample in the target domain, that is, use the first speech recognition model to perform speech recognition on each unlabeled audio sample to obtain the first pseudo-label of each unlabeled audio sample.

[0098] Step 4: Supervised model pre-training is performed on N initial second models using multiple second labeled audio samples in the source domain to obtain N second speech recognition models;

[0099] Step 5: For each of the N second speech recognition models, decode each unlabeled audio sample using the second speech recognition model, that is, perform speech recognition on each unlabeled audio sample using the second speech recognition model, thereby obtaining multiple second pseudo-labels for each unlabeled audio sample corresponding to each second speech recognition model.

[0100] Step 6: Select multiple target pseudo-labels from the multiple second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model;

[0101] Step 7: Using multiple target pseudo-labels and the corresponding unlabeled audio samples for each target pseudo-label, perform supervised fine-tuning training on the pre-trained model to obtain the target speech recognition model.

[0102] It should be noted that, among steps ① to ⑦ above, step ④ can be performed after step ③ (e.g., Figure 4 (As shown), it can also be performed before step ① above, or it can be performed simultaneously with step ① above. This disclosure does not make any specific limitations on this.

[0103] In addition, this disclosure also provides a speech recognition method, which includes:

[0104] Obtain the audio data to be recognized;

[0105] Audio data is input into a pre-trained target speech recognition model to obtain the target text corresponding to the target audio data. The target speech recognition model is trained according to the speech recognition model generation method provided in this disclosure.

[0106] Figure 5 This is a block diagram illustrating a speech recognition model generation apparatus according to an exemplary embodiment. Figure 5 As shown, the device 200 includes:

[0107] The first acquisition module 201 is used to acquire multiple first labeled audio samples and multiple unlabeled audio samples within the target domain;

[0108] The first pre-training module 202 is used to perform self-supervised model pre-training on the first initial model using the multiple unlabeled audio samples to obtain a pre-trained model.

[0109] The generation module 203 is used to generate a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model;

[0110] The fine-tuning module 204 is used to perform supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples to obtain the target speech recognition model.

[0111] In the above technical solution, the model is first pre-trained using unlabeled audio samples within the target domain. Then, pseudo-labels for the unlabeled audio samples within the target domain are generated using labeled audio samples and the pre-trained model. Finally, the pre-trained model is fine-tuned using the unlabeled audio samples and the corresponding pseudo-labels to obtain the target speech recognition model. This allows pseudo-labels to compensate for the low accuracy of the speech recognition model caused by insufficient labeled audio samples in the target domain, thereby improving the accuracy of the speech recognition model. Furthermore, the entire model generation stage uses training data within the target domain, avoiding interference from general domain training data on the speech recognition model's ability in the target domain, thus ensuring the target speech recognition model's speech recognition capability within the target domain. In addition, the accuracy of the speech recognition model can be guaranteed with only a small number of labeled audio samples, eliminating the need for extensive manual annotation, saving manpower and improving the training efficiency of the speech recognition model.

[0112] Optionally, the generation module 203 includes:

[0113] The first fine-tuning submodule is used to perform supervised fine-tuning training on the pre-trained model using the multiple first labeled audio samples to obtain a first speech recognition model.

[0114] The decoding submodule is used to perform speech recognition on each unlabeled audio sample using the first speech recognition model to obtain the first pseudo-label of the unlabeled audio sample.

[0115] Optionally, the device 200 further includes:

[0116] The third acquisition module is used to acquire multiple second labeled audio samples in the source domain before the fine-tuning module 204 performs supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the multiple unlabeled audio samples.

[0117] The second pre-training module is used to perform supervised model pre-training on N second initial models using the multiple second labeled audio samples to obtain N second speech recognition models, where N≥1;

[0118] The decoding module is used to perform speech recognition on each of the N second speech recognition models for each of the unlabeled audio samples, thereby obtaining multiple second pseudo-labels for the unlabeled audio samples corresponding to the second speech recognition model.

[0119] The fine-tuning module 204 is used to perform supervised fine-tuning training on the pre-trained model using each first pseudo-label, each second pseudo-label corresponding to the second speech recognition model, and the multiple unlabeled audio samples, to obtain the target speech recognition model.

[0120] Optionally, the fine-tuning module 204 includes:

[0121] The filtering submodule is used to filter out multiple target pseudo-labels from the plurality of second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model;

[0122] The second fine-tuning submodule is used to perform supervised fine-tuning training on the pre-trained model using the multiple target pseudo-labels and the unlabeled audio samples corresponding to each target pseudo-label, so as to obtain the target speech recognition model.

[0123] Optionally, N = 1;

[0124] The filtering submodule includes:

[0125] The first calculation submodule is used to calculate the first similarity between the first pseudo-label and the second pseudo-label of each unlabeled audio sample.

[0126] The first determining submodule is used to determine a target pseudo-label from the first pseudo-label and the second pseudo-label of the unlabeled audio sample if the first similarity is greater than the first preset similarity threshold.

[0127] Optionally, the first determining submodule is used to randomly select one of the first pseudo-label and the second pseudo-label of the unlabeled audio sample as the target pseudo-label.

[0128] Optionally, the first determining submodule is used to determine the pseudo-label generated by a speech recognition model with a low word error rate from the first pseudo-label and the second pseudo-label of the unlabeled audio sample as the target pseudo-label.

[0129] Optionally, N>1;

[0130] The filtering submodule includes:

[0131] The second calculation submodule is used to calculate the second similarity between each pair of the first pseudo-label and N second pseudo-labels of the unlabeled audio sample for each unlabeled audio sample.

[0132] The second determining submodule is used to determine the target pseudo-label from the two pseudo-labels corresponding to the maximum value of the second similarity between the pairs of the unlabeled audio sample if the maximum value is greater than the second preset similarity threshold.

[0133] Optionally, N > 1, and the N second initial models are all different from each other.

[0134] Figure 6 This is a block diagram illustrating a speech recognition device according to an exemplary embodiment. Figure 6 As shown, the device 300 includes:

[0135] The second acquisition module 301 is used to acquire the audio data to be recognized;

[0136] The speech recognition module 302 is used to input the audio data into a pre-trained target speech recognition model to obtain the target text corresponding to the target audio data, wherein the target speech recognition model is trained according to the speech recognition model generation method provided in this disclosure.

[0137] The following is for reference. Figure 7 The diagram illustrates a structural schematic of an electronic device (e.g., a terminal device or a server) 600 suitable for implementing embodiments of the present disclosure. The terminal device in the embodiments of the present disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0138] like Figure 7 As shown, electronic device 600 may include a processing device (e.g., a central processing unit, a graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage device 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of electronic device 600. Processing device 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0139] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 An electronic device 600 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0140] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, it performs the functions defined in the methods of embodiments of this disclosure.

[0141] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0142] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0143] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0144] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: acquire multiple first labeled audio samples and multiple unlabeled audio samples within a target domain; perform self-supervised model pre-training on a first initial model using the multiple unlabeled audio samples to obtain a pre-trained model; generate a first pseudo-label for each of the unlabeled audio samples using the multiple first labeled audio samples and the pre-trained model; and perform supervised fine-tuning training on the pre-trained model using each first pseudo-label and the multiple unlabeled audio samples to obtain a target speech recognition model.

[0145] Alternatively, the aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire audio data to be recognized; input the audio data into a pre-trained target speech recognition model to obtain target text corresponding to the target audio data, wherein the target speech recognition model is trained by the speech recognition model generation method provided in this disclosure.

[0146] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including but not limited to object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0147] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0148] The modules described in the embodiments of this disclosure can be implemented in software or in hardware. The names of the modules are not necessarily limiting in certain circumstances; for example, the first acquisition module can also be described as "a module for acquiring multiple first labeled audio samples and multiple unlabeled audio samples within a target domain".

[0149] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0150] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0151] According to one or more embodiments of this disclosure, Example 1 provides a speech recognition model generation method, comprising: acquiring a plurality of first labeled audio samples and a plurality of unlabeled audio samples in a target domain; performing self-supervised model pre-training on a first initial model using the plurality of unlabeled audio samples to obtain a pre-trained model; generating a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model; and performing supervised fine-tuning training on the pre-trained model using each first pseudo-label and the plurality of unlabeled audio samples to obtain a target speech recognition model.

[0152] According to one or more embodiments of this disclosure, Example 2 provides the method of Example 1, wherein generating a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model includes: performing supervised fine-tuning training on the pre-trained model using the plurality of first labeled audio samples to obtain a first speech recognition model; and performing speech recognition on each of the unlabeled audio samples using the first speech recognition model to obtain a first pseudo-label for the unlabeled audio sample.

[0153] According to one or more embodiments of this disclosure, Example 3 provides the method of Example 2, which, before the step of performing supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples, further includes: acquiring a plurality of second labeled audio samples in the source domain; performing supervised model pre-training on N second initial models using the plurality of second labeled audio samples to obtain N second speech recognition models, where N≥1; for each of the N second speech recognition models, performing speech recognition on each of the unlabeled audio samples using the second speech recognition model to obtain a plurality of second pseudo-labels corresponding to the second speech recognition model and the unlabeled audio samples; the step of performing supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples to obtain a target speech recognition model includes: performing supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels, the plurality of second pseudo-labels corresponding to each of the second speech recognition models, and the plurality of unlabeled audio samples to obtain the target speech recognition model.

[0154] According to one or more embodiments of this disclosure, Example 4 provides the method of Example 3, wherein the method of performing supervised fine-tuning training on the pre-trained model using each first pseudo-label, the plurality of second pseudo-labels corresponding to each second speech recognition model, and the plurality of unlabeled audio samples to obtain a target speech recognition model includes: selecting a plurality of target pseudo-labels from each first pseudo-label and the plurality of second pseudo-labels corresponding to each second speech recognition model; and performing supervised fine-tuning training on the pre-trained model using the plurality of target pseudo-labels and the unlabeled audio samples corresponding to each target pseudo-label to obtain a target speech recognition model.

[0155] According to one or more embodiments of this disclosure, Example 5 provides the method of Example 4, where N=1; the step of selecting multiple target pseudo-labels from the plurality of second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model includes: for each unlabeled audio sample, calculating a first similarity between the first pseudo-label and the second pseudo-label of the unlabeled audio sample; if the first similarity is greater than a first preset similarity threshold, then determining a target pseudo-label from the first pseudo-label and the second pseudo-label of the unlabeled audio sample.

[0156] According to one or more embodiments of this disclosure, Example 6 provides the method of Example 5, wherein determining the target pseudo-tag from the first pseudo-tag and the second pseudo-tag of the unlabeled audio sample includes: randomly selecting one of the first pseudo-tag and the second pseudo-tag of the unlabeled audio sample as the target pseudo-tag.

[0157] According to one or more embodiments of this disclosure, Example 7 provides the method of Example 5, wherein determining the target pseudo-label from the first pseudo-label and the second pseudo-label of the unlabeled audio sample includes: determining the pseudo-label generated by a speech recognition model with a low word error rate from the first pseudo-label and the second pseudo-label of the unlabeled audio sample as the target pseudo-label.

[0158] According to one or more embodiments of this disclosure, Example 8 provides the method of Example 4, where N>1; the step of selecting multiple target pseudo-labels from the plurality of second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model includes: for each unlabeled audio sample, calculating the second similarity between each pair of the first pseudo-label and N second pseudo-labels of the unlabeled audio sample; if the maximum value of the second similarity between each pair of the unlabeled audio sample is greater than a second preset similarity threshold, then determining the target pseudo-label from the two pseudo-labels corresponding to the maximum value.

[0159] According to one or more embodiments of this disclosure, Example 9 provides the method of any one of Examples 3, 4, and 8, where N > 1 and the N second initial models are different from each other.

[0160] According to one or more embodiments of this disclosure, Example 10 provides a speech recognition method, the method comprising: acquiring audio data to be recognized; inputting the audio data into a pre-trained target speech recognition model to obtain target text corresponding to the target audio data, wherein the target speech recognition model is trained according to the speech recognition model generation method of any one of Examples 1-9.

[0161] According to one or more embodiments of this disclosure, Example 11 provides a speech recognition model generation apparatus, comprising: a first acquisition module, configured to acquire a plurality of first labeled audio samples and a plurality of unlabeled audio samples in a target domain; a first pre-training module, configured to perform self-supervised model pre-training on a first initial model using the plurality of unlabeled audio samples to obtain a pre-trained model; a generation module, configured to generate a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model; and a fine-tuning module, configured to perform supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples to obtain a target speech recognition model.

[0162] According to one or more embodiments of this disclosure, Example 12 provides an apparatus of Example 11, wherein the generation module includes: a first fine-tuning submodule, configured to perform supervised fine-tuning training on the pre-trained model using the plurality of first labeled audio samples to obtain a first speech recognition model; and a decoding submodule, configured to perform speech recognition on each of the unlabeled audio samples using the first speech recognition model to obtain a first pseudo-label for the unlabeled audio sample.

[0163] According to one or more embodiments of this disclosure, Example 13 provides an apparatus of Example 12, the apparatus further comprising: a third acquisition module, configured to acquire a plurality of second labeled audio samples in the source domain before the fine-tuning module performs supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples; a second pre-training module, configured to perform supervised model pre-training on N second initial models using the plurality of second labeled audio samples to obtain N second speech recognition models, wherein N≥1; a decoding module, configured to perform speech recognition on each of the unlabeled audio samples using each of the N second speech recognition models to obtain a plurality of second pseudo-labels corresponding to the second speech recognition model and the unlabeled audio samples; the fine-tuning module is configured to perform supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels, the plurality of second pseudo-labels corresponding to each of the second speech recognition models, and the plurality of unlabeled audio samples to obtain a target speech recognition model.

[0164] According to one or more embodiments of this disclosure, Example 14 provides an apparatus of Example 13, wherein the fine-tuning module includes: a screening submodule for screening a plurality of target pseudo-labels from the plurality of second pseudo-labels corresponding to each first pseudo-label and each second speech recognition model; and a second fine-tuning submodule for performing supervised fine-tuning training on the pre-trained model using the plurality of target pseudo-labels and the unlabeled audio samples corresponding to each target pseudo-label to obtain a target speech recognition model.

[0165] According to one or more embodiments of this disclosure, Example 15 provides the apparatus of Example 14, where N=1; the filtering submodule includes: a first calculation submodule, configured to calculate a first similarity between the first pseudo-label and the second pseudo-label of each unlabeled audio sample; and a first determination submodule, configured to determine a target pseudo-label from the first pseudo-label and the second pseudo-label of the unlabeled audio sample if the first similarity is greater than a first preset similarity threshold.

[0166] According to one or more embodiments of this disclosure, Example 16 provides the apparatus of Example 15, wherein the first determining submodule is configured to randomly select one of the first pseudo-tag and the second pseudo-tag of the unlabeled audio sample as a target pseudo-tag.

[0167] According to one or more embodiments of this disclosure, Example 17 provides the apparatus of Example 15, wherein the first determining submodule is configured to determine the pseudo-label generated by a speech recognition model with a low word error rate from the first pseudo-label and the second pseudo-label of the unlabeled audio sample as the target pseudo-label.

[0168] According to one or more embodiments of this disclosure, Example 18 provides the apparatus of Example 14, where N>1; the filtering submodule includes: a second calculation submodule, configured to calculate, for each unlabeled audio sample, a second similarity between each pair of the first pseudo-label and N second pseudo-labels of the unlabeled audio sample; and a second determination submodule, configured to determine a target pseudo-label from the two pseudo-labels corresponding to the maximum value if the maximum value of the second similarity between each pair of the unlabeled audio sample is greater than a second preset similarity threshold.

[0169] According to one or more embodiments of this disclosure, Example 19 provides an apparatus as described in any one of Examples 13, 14, and 18, where N > 1, and the N second initial models are distinct from each other.

[0170] According to one or more embodiments of this disclosure, Example 20 provides a speech recognition device, the device comprising: a second acquisition module for acquiring audio data to be recognized; and a speech recognition module for inputting the audio data into a pre-trained target speech recognition model to obtain target text corresponding to the target audio data, wherein the target speech recognition model is trained according to the speech recognition model generation method described in any one of Examples 1-9.

[0171] According to one or more embodiments of the present disclosure, Example 21 provides a computer-readable medium having a computer program stored thereon that, when executed by a processing device, implements the steps of the method described in any one of Examples 1-10.

[0172] According to one or more embodiments of this disclosure, Example 22 provides an electronic device, including:

[0173] A storage device having at least one computer program stored thereon;

[0174] At least one processing means is configured to execute the at least one computer program in the storage device to implement the steps of the method described in any one of Examples 1-10.

[0175] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0176] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0177] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative forms of implementing the claims. Regarding the apparatus in the above embodiments, the specific manner in which the various modules perform their operations has been described in detail in the embodiments relating to the method, and will not be elaborated upon here.

Claims

1. A method for generating a speech recognition model, characterized in that, include: Obtain multiple first-labeled audio samples and multiple unlabeled audio samples within the target domain; The first initial model is pre-trained using the multiple unlabeled audio samples to obtain a pre-trained model. Using the plurality of first labeled audio samples and the pre-trained model, a first pseudo-label is generated for each of the unlabeled audio samples; The pre-trained model is subjected to supervised fine-tuning training using each of the first pseudo-labels and the plurality of unlabeled audio samples to obtain the target speech recognition model. The step of generating a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model includes: The pre-trained model is trained in a supervised manner using the multiple first labeled audio samples to obtain a first speech recognition model. For each unlabeled audio sample, the first speech recognition model is used to perform speech recognition on the unlabeled audio sample to obtain the first pseudo-label of the unlabeled audio sample. Prior to the step of performing supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples, the method further includes: Obtain multiple second-labeled audio samples within the source domain; Using the multiple second labeled audio samples, supervised model pre-training is performed on N second initial models to obtain N second speech recognition models, where N≥1; For each of the N second speech recognition models, speech recognition is performed on each of the unlabeled audio samples using the second speech recognition model to obtain multiple second pseudo-labels for the unlabeled audio samples corresponding to the second speech recognition model. The step of using each of the first pseudo-labels and the plurality of unlabeled audio samples to perform supervised fine-tuning training on the pre-trained model to obtain the target speech recognition model includes: The pre-trained model is supervisedly fine-tuned using each of the first pseudo-labels, each of the second pseudo-labels corresponding to the second speech recognition model, and the multiple unlabeled audio samples to obtain the target speech recognition model.

2. The method according to claim 1, characterized in that, The step of using each of the first pseudo-labels, the plurality of second pseudo-labels corresponding to each of the second speech recognition models, and the plurality of unlabeled audio samples to perform supervised fine-tuning training on the pre-trained model to obtain the target speech recognition model includes: Multiple target pseudo-labels are selected from the plurality of second pseudo-labels corresponding to each of the first pseudo-labels and each of the second speech recognition models; The pre-trained model is subjected to supervised fine-tuning training using the multiple target pseudo-labels and the unlabeled audio samples corresponding to each target pseudo-label to obtain the target speech recognition model.

3. The method according to claim 2, characterized in that, N=1； The step of selecting multiple target pseudo-labels from the plurality of second pseudo-labels corresponding to each of the first pseudo-labels and each of the second speech recognition models includes: For each unlabeled audio sample, calculate the first similarity between the first pseudo-label and the second pseudo-label of the unlabeled audio sample; If the first similarity is greater than the first preset similarity threshold, then the target pseudo-label is determined from the first pseudo-label and the second pseudo-label of the unlabeled audio sample.

4. The method according to claim 3, characterized in that, Determining the target pseudo-tag from the first pseudo-tag and the second pseudo-tag of the unlabeled audio sample includes: One of the first and second pseudo-labels of the unlabeled audio sample is randomly selected as the target pseudo-label.

5. The method according to claim 3, characterized in that, Determining the target pseudo-tag from the first pseudo-tag and the second pseudo-tag of the unlabeled audio sample includes: The pseudo-label generated by the speech recognition model with a low word error rate from the first pseudo-label and the second pseudo-label of the unlabeled audio sample is determined as the target pseudo-label.

6. The method according to claim 2, characterized in that, N>1； The step of selecting multiple target pseudo-labels from the plurality of second pseudo-labels corresponding to each of the first pseudo-labels and each of the second speech recognition models includes: For each unlabeled audio sample, calculate the second similarity between each pair of the first pseudo-label and N second pseudo-labels of the unlabeled audio sample; If the maximum value of the second similarity between any two pairs of the unlabeled audio sample is greater than the second preset similarity threshold, then the target pseudo-label is determined from the two pseudo-labels corresponding to the maximum value.

7. The method according to any one of claims 1, 2, and 6, characterized in that, N > 1, and the N second initial models are all different from each other.

8. A speech recognition method, characterized in that, The method includes: Obtain the audio data to be recognized; The audio data is input into a pre-trained target speech recognition model to obtain the target text corresponding to the target audio data, wherein the target speech recognition model is trained by the speech recognition model generation method according to any one of claims 1-7.

9. A speech recognition model generation device, characterized in that, include: The first acquisition module is used to acquire multiple first labeled audio samples and multiple unlabeled audio samples within the target domain; The first pre-training module is used to perform self-supervised model pre-training on the first initial model using the multiple unlabeled audio samples to obtain a pre-trained model. The generation module is used to generate a first pseudo-label for each of the unlabeled audio samples using the plurality of first labeled audio samples and the pre-trained model; The fine-tuning module is used to perform supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the plurality of unlabeled audio samples to obtain the target speech recognition model. The generation module includes: The first fine-tuning submodule is used to perform supervised fine-tuning training on the pre-trained model using the multiple first labeled audio samples to obtain a first speech recognition model. The decoding submodule is used to perform speech recognition on each unlabeled audio sample using the first speech recognition model to obtain the first pseudo-label of the unlabeled audio sample. The device further includes: The third acquisition module is used to acquire multiple second labeled audio samples in the source domain before the fine-tuning module performs supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels and the multiple unlabeled audio samples. The second pre-training module is used to perform supervised model pre-training on N second initial models using the multiple second labeled audio samples to obtain N second speech recognition models, where N≥1; The decoding module is used to perform speech recognition on each of the N second speech recognition models for each of the unlabeled audio samples, thereby obtaining multiple second pseudo-labels for the unlabeled audio samples corresponding to the second speech recognition model. The fine-tuning module is used to perform supervised fine-tuning training on the pre-trained model using each of the first pseudo-labels, the plurality of second pseudo-labels corresponding to each of the second speech recognition models, and the plurality of unlabeled audio samples, to obtain the target speech recognition model.

10. A voice recognition device, characterized in that, The device includes: The second acquisition module is used to acquire the audio data to be recognized; A speech recognition module is used to input the audio data into a pre-trained target speech recognition model to obtain the target text corresponding to the target audio data, wherein the target speech recognition model is trained by the speech recognition model generation method according to any one of claims 1-7.

11. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processing device, it implements the steps of the method described in any one of claims 1-8.

12. An electronic device, characterized in that, include: A storage device having at least one computer program stored thereon; At least one processing means is configured to execute the at least one computer program in the storage device to implement the steps of the method according to any one of claims 1-8.

Images